Neural network based diagnosis of mechanical faults in IC engines

Abstract

Oversize clearance induced piston slap and big end bearing knock are two common mechanical faults in the operation of internal combustion (IC) engines. A previous study has shown that the vibration signals measured on the engine block can be used to diagnose such mechanical faults in engines. However this requires some advanced signal processing techniques to be applied. Envelope analysis converts the signals from piston slap and bearing knock (second order cyclostationary signals) into deterministic signals, to which synchronous averaging can be applied. Before generating the envelope, the “kurtogram” was used to filter the signal and find the frequency bands with high impulsiveness. The amplitudes and phases of the Fourier series of the averaged envelope signals were extracted as diagnostic features. In order to realize automated and intelligent fault diagnosis of the engine, Artificial Neural Networks (ANN) were trained using the features characteristic of different faults. The networks comprise three stages: fault detection, fault localization and severity identification. The critical issue for successful application of ANNs in fault diagnosis is the selection of optimal features from the candidate ones. Two feature selection methods – “filter” and “wrapper”- were used to select the optimal features for the fault detection. The Relief approach is a typical “filter” feature selection method and Genetic Algorithm (GA) is a typical “wrapper” method. The features selected from the two approaches were separately used as the inputs to ANNs and the results are compared. Because the “wrapper” method takes into account the relevance of the individual features, the comparison showed that the GA approach has advantages in feature selection for fault diagnosis.